Google has published a study examining the environmental footprint of artificial intelligence across its infrastructure. According to the report, a typical text prompt to its Gemini AI model consumes less energy than watching nine seconds of television and requires just 0.26 milliliters of water, about the equivalent of five drops.
These findings appear to position AI queries as relatively low-impact activities, suggesting that fears about AI’s environmental burden may be overstated. However, researchers cited in the report contend that the figures do not provide a complete picture and may leave the public with a misleading impression.
Experts Call Out Gaps in Google’s Approach
One of the study’s listed authors, Shaolei Ren, an associate professor at the University of California, told The Verge that Google’s framing avoids key details. His criticism centers on the exclusion of indirect water usage—the water consumed at power plants that supply electricity to data centers.
The water figure presented by Google only reflects cooling systems inside its facilities. But as the International Energy Agency (IEA) has pointed out, roughly 60% of water associated with data centers comes from external sources, primarily electricity generation. This includes water used to produce steam and cool turbines in fossil-fuel power plants.
By leaving this out, experts argue, Google significantly underrepresents the true scope of AI’s water demands.
Direct vs. Indirect Water Consumption
The contrast between Google’s numbers and independent research is striking. Ren’s earlier work suggested that a single AI query could use around 50 milliliters of water—almost 200 times higher than Google’s figure. His estimates included indirect consumption, while Google’s did not.
The debate is particularly relevant because many large data centers are being constructed in water-scarce regions. Even tiny per-query amounts can accumulate into substantial resource strain when scaled to billions of global queries. For local communities already grappling with water shortages, the impact could be significant.
Training Models: The Hidden Resource Drain
Another omission in Google’s report is the cost of training massive AI models. While everyday usage figures are highlighted, training requires immense computing power and prolonged operation of thousands of GPUs, leading to vastly higher consumption of both water and energy.
Although Google has not disclosed its own figures, other companies have. French AI startup Mistral reported that training its Large 2 model generated about 20.4 kilotons of carbon emissions and consumed 281,000 cubic meters of water—comparable to filling 112 Olympic-sized swimming pools.
By focusing narrowly on usage rather than training, Google leaves out a critical part of AI’s environmental story.
Carbon Emissions: Market vs. Location-Based Accounting
The study has also drawn criticism for how it measures carbon emissions. Experts note that Google relied on a market-based approach, which accounts for renewable energy credits purchased by utilities rather than the actual electricity mix powering data centers in specific regions.
This method allows emissions to appear lower than they are in practice. A location-based analysis, on the other hand, would measure the precise share of renewable and fossil-fuel energy feeding each facility. This approach often reveals higher carbon impacts, especially for data centers in areas heavily reliant on coal or natural gas.
Researcher Alex de Vries-Gao, whose work was cited in the report, has emphasized the importance of location-specific data to accurately capture AI’s carbon footprint.
Median vs. Average Calculations
Beyond omissions, the methodology of Google’s study has raised concerns. The company reported results using a median prompt size, while previous studies, including Ren’s, used averages. Without transparency about how Google calculated the median, comparisons are difficult.
Additionally, the study only included text-based queries, excluding far more resource-intensive tasks like image and video generation, which Gemini is capable of performing. This omission further narrows the scope of the results.
Other AI Firms Make Similar Claims
Google is not the only company putting forward small-sounding usage figures. Earlier this year, OpenAI CEO Sam Altman said a single ChatGPT query consumes about one-fifteenth of a teaspoon of water. Like Google’s number, this sounds trivial at face value. But when scaled to billions of queries daily, these figures translate into considerable environmental demand.
Experts caution that such per-query statistics oversimplify the issue and risk obscuring the much larger question of how AI infrastructure expansion affects global resources.
The Bigger Challenge: Scaling Up AI
While efficiency gains per query are notable, the rapid growth of AI infrastructure is intensifying overall consumption. Data centers are multiplying worldwide, creating huge new demands for power and water.
Analysts warn that this expansion is already straining energy grids and local water supplies, especially in regions where these resources are limited. Improvements in efficiency may reduce the impact at the margins, but the exponential scale of AI adoption risks negating those benefits.
The controversy surrounding Google’s report underscores the need for standardized reporting methods to evaluate AI’s environmental impact. Without clear and consistent frameworks, companies can present selective figures that highlight efficiency gains while minimizing or omitting less flattering data.
Experts argue that credible assessments must include direct and indirect water consumption, location-specific energy mixes, and the costs of model training, alongside per-query efficiency measures. Until such standards are adopted, comparisons across companies and models will remain inconsistent and incomplete.
